Two systems exist so far: a prosthesis grip controller and a Japanese-to-English translation pipeline. Both are latency-critical machine-learning systems that run locally, on hardware the operator controls.
Sole proprietorship · Bayreuth, Germany · trading since 2026
ayaka is the trade name of a one-person software and machine-learning business, registered in Bayreuth on 11 August 2026. It builds systems whose inference runs on the operator's own machines, publishes its research as open source, and reports the measurements that failed next to the ones that worked.
Prototype · public demo
A 2.23-million-parameter model beats a detector that is handed the right answer
A wrist camera shapes the grip before contact. The muscle signal decides when to close; the vision system decides how.
72.3 % grip accuracy · 2.23M parameters · pixels only · Apache-2.0
Active research
Four small models per paragraph, and one result that had to be withdrawn
Long-form literary Japanese into English, drafted, merged, polished and reviewed without a page leaving the machine.
four-stage pipeline · single consumer GPU · fully offline
Local-first
Inference that has to answer in 125 milliseconds cannot wait for a network.
Local-first here is a constraint, not a preference. A prosthetic grip has to be committed before the hand reaches the object, which rules out a round trip to a rented endpoint. A translation pipeline that reads an entire book has no business shipping that book to a third party. Both problems end at the same place: the model runs on the machine in front of you.
What follows from it: models are trained on hardware here, they ship in formats that run without a vendor, and the numbers that describe them are measured on the same hardware they run on.
Taken from the trade description registered on 11 August 2026. Work outside it is declined rather than improvised.
01
Development, adaptation and operation, from a single-purpose internal tool to a browser application that has to run offline.
02
Development, training and licensing of models, and the services built on them — machine translation and automated document capture.
03
Technical consulting, adaptation, maintenance and operation of systems that are already running.
04
Creation and editing of graphics, video and audio — the material a technical project needs in order to be understood.
How the work is done
The failed measurements are published next to the ones that worked.
A model that only ever produced good numbers has not been measured hard enough. The Hime evaluation once produced a positive result that turned out to come from a leak in the answer key; it was traced, withdrawn, and the harness rebuilt around it. That is on the project page, not buried in a changelog.
Client material stays out of third-party endpoints. Training happens on hardware here. Every figure quoted on this site is one that was measured, and where a number would be convenient but is not known, the sentence is written without it.